UCL Discovery

Abstract

Automatic reconstruction of fragmented objects is of great interest in archaeology, where artefacts are often found in a fractured state. In this paper, we focus on the problem of automatically agglomerating clusters of fragments from previously determined pairwise matches. Common to any automated cluster agglomeration technique is the challenge of error accumulation, making it increasingly difficult to discern false from true matches as the assembly grows. Many assembly algorithms therefore introduce a global relaxation phase to distribute alignment errors evenly across the cluster, minimising major inconsistencies. Nevertheless, error accumulation limits the problem size automated assembly systems can handle in practice. In this paper we show how two careful modifications of the traditional relaxation scheme help lift this limit considerably. In contrast to previous work, we integrate global relaxation earlier, in the search phase of the assembly process. In addition, we do not fix connections between assembled fragments, but rather leave them flexible throughout the assembly. By modifying two representative assembly algorithms, we demonstrate the effectiveness of our approach. Using the example of a challenging fresco dataset, we show that these modifications achieve larger reconstruction sizes than traditional strategies.